2026 AI Whitepaper Glossary

Agent (AI agent)

An AI agent is an autonomous software entity that can perceive information from its environment (e.g., data, documents, APIs), maintain state, and select actions to achieve objectives. In modern systems, an agent often combines a reasoning or policy component with tool use (search, retrieval, code execution) and can operate over multiple steps rather than producing a single response.

Agentic AI

Agentic AI refers to AI systems designed to pursue goals through iterative planning and action: they decompose tasks, choose among actions or tools, observe outcomes, and adapt until a stopping condition is reached. Agentic systems may orchestrate multiple models or modules (e.g., retrieval, parsing, validation) within a closed-loop workflow.

Agentic misalignment

Agentic misalignment is the failure mode in which an autonomous system’s optimized objective (e.g. an explicit reward) diverges from the designer’s actual intended goal, producing undesired actions. It includes cases where the system follows the given objective in unintended ways because the goal was incomplete, ambiguous, or insufficiently constrained.

AI literacy

AI literacy is the organizational capability to use AI correctly and safely: understanding what models can/cannot do, how errors arise (e.g., hallucination), how to validate outputs, and how to apply governance. AI literacy is a critical governance and risk management component, by ensuring staff can challenge and verify AI outputs, understand and identify model limitations.

Alignment methods
Alignment methods are techniques used to steer AI systems toward behaviours and outputs that are consistent with human intentions, making sure AI systems behave in the way people want them to: avoiding harmful behaviour, following instructions, and producing useful results rather than just optimizing raw performance. These methods include data curation, reward modelling, reinforcement learning from human feedback (RLHF), rule-based constraints, safety filters, and post-training fine-tuning.

Anomaly detection

Anomaly detection is the task of identifying observations that deviate markedly from expected patterns in data. Methods may be statistical, distance-based, density-based, or model-based, and are often used when anomalies are rare and labels are unavailable or incomplete.

Auditability (of AI)

Auditability is the extent to which an AI system’s outputs and decisions can be reconstructed, inspected, and verified after the fact. It typically requires traceability of inputs, data transformations, model/version identifiers, prompts or configurations, and logs of intermediate steps so that results are reproducible and reviewable.

Backpropagation

Backpropagation is an algorithm for efficiently computing gradients of a loss function with respect to the parameters of a neural network by applying the chain rule through the network’s computational graph. These gradients are then used by optimization methods (e.g., stochastic gradient descent) to update parameters during training. In other words, it is the method by which neural networks learn from mistakes: figuring out which internal settings caused an error and adjusting them step by step.

Bayesian imputation

Bayesian imputation is the estimation of missing values by specifying a probabilistic model for the data and inferring a posterior distribution over the missing entries. Unlike single-point imputations, it represents uncertainty and can generate multiple plausible imputations consistent with observed data. Simply put, instead of guessing a single missing value, this approach estimates a range of likely values and how uncertain that estimate is.

Bias (in AI context)

Model bias is a systematic tendency of a model to produce errors in a particular direction or to produce systematically different outcomes across groups due to data, measurement, labelling, or modelling choices. It occurs when an AI system consistently favours or disadvantages certain cases because of how it was trained or designed.

Carbon footprint of a portfolio

A carbon footprint refers to the entire greenhouse gas (GHG) emissions of a portfolio. It is calculated in tons of CO2 equivalents per million USD invested (tCO2e/mUSD). It expresses the amount of annual GHG emissions which can be allocated to the investor per million USD invested in a portfolio and is therefore probably the most intuitive carbon metric available at the portfolio level.

Carbon Credits / Carbon Markets         

Carbon markets are trading schemes in which participants buy and sell "carbon credits" that allow them to emit a certain amount of CO2 or other greenhouse gases (GHG). The primary goal of carbon markets is to reduce overall GHG emissions by creating a monetary incentive for participants to reduce their carbon emissions.

There are two types of carbon markets:

• Regulated carbon markets (also called "Emissions Trading Systems" or "cap-and-trade systems"), whereby a government or regulatory body sets a cap on total emissions and allocates or sells emission allowances to participants. Participation is required by law. If a participant emits less than its allowance, it can sell the excess to others who need more. Credits traded on these markets are typically called "compliance carbon credits". Over time, the cap is usually reduced, encouraging overall emission reductions. Examples include the European Union Emissions Trading System (EU ETS) or California's Cap-and-Trade Program.

• Voluntary carbon markets are non-regulated markets where participation is optional. Participants can buy voluntary carbon credits generated by projects that reduce or remove emissions (such as reforestation or renewable energy projects), with the purpose to compensate for their own emissions. Given the additionality of carbon reduction generated by related projects can often not be guaranteed, it cannot be expected that own emissions are fully offset by the purchase of carbon credits. Nevertheless, voluntary carbon markets contribute to the reduction of CO2-emissions if projects that would not be economically viable without the revenues of the carbon credits become profitable.

Classification

Classification is a supervised learning task in which a model assigns inputs to one of a predefined set of categories or labels based on learned patterns in labelled training data. It corresponds to an AI system deciding which “bucket” something belongs to, such as labelling an email as spam or not spam, or identifying whether a text talks about one topic or another.

Climate risk

Climate risk refers to the potential impact of climate change to human societies coming from climate change and its related effects - extreme heat waves, droughts, wildfires and floods. It includes impact on lives, health and wellbeing, economic and social structures, cultural assets, infrastructure, financial investments, etc.

There are two categories of climate risk: Physical risks that arise from the physical impacts of climate change (e.g. assets affected by storms or floods) and transition risks related to the adaptation of the political frameworks and economic system (e.g. costs resulting from climate policy measures).

Climate risk is directly relevant to the financial services sector because it has the potential to impact asset values, investment returns, and the financial stability and resilience of investee companies or borrowers.

Clustering

Clustering is an unsupervised learning task that groups data points into clusters such that points within the same cluster are more similar to each other than to points in other clusters, according to a chosen similarity or distance measure. Clustering helps discover natural groupings in data, like grouping documents by topic or customers by behavior, without being told what those groups should be.

Computer vision

Computer vision allows computers to “see”, for example, identifying objects in photos or detecting changes in satellite images. It is the field of AI focused on enabling machines to interpret and act on visual data (images and video). Core tasks include classification, detection, segmentation, tracking, and estimating properties of scenes and objects.

Convolutional Neural Network

A convolutional neural network (CNN) is a neural architecture designed for grid-structured data that uses convolutional filters to learn local patterns and hierarchical features. Weight sharing and locality make CNNs parameter-efficient and effective for translation-tolerant feature extraction. CNNs are especially good at understanding images because they look at small visual patterns first and then build up to bigger shapes and structures.

Decarbonisation / Carbon Neutrality

Carbon neutrality refers to the situation when an organisation’s net carbon emissions are equal to zero. The process requires measuring total CO2 emissions, taking active steps to reduce emissions where the company can, and then purchasing carbon credits to offset CO2 emissions that cannot be eliminated from a company's operations. The carbon credits contribute to financing projects reducing CO2-emissions (i.e. by replacing fossil power generation with renewable energy projects).

Decision tree

A decision tree is a supervised learning model that predicts an outcome by recursively partitioning the feature space using rule-like splits. The model asks a series of yes/no questions until it reaches a conclusion. Trees can be used for classification or regression and are valued for interpretability, though they can overfit without regularization or pruning.

Deep learning

Deep learning is a subset of machine learning that uses neural networks with many layers to learn representations and predictive functions from data. It is particularly effective for unstructured inputs such as text, audio, and images.

Distribution shift

Distribution shift occurs when the statistical properties of data encountered during deployment differ from those present in the training data, in other words when the world changes but the model does not, causing predictions to become unreliable over time. Common forms include covariate shift, label shift, and concept drift, and shift is a leading cause of performance degradation in production.

Distributional robustness

Distributional robustness refers to modeling and evaluation approaches that aim to maintain acceptable performance when the test distribution differs from the training distribution, including under worst-case or adversarial perturbations. Techniques include robust optimization, stress testing, and uncertainty-aware decision rules.

Domain adaptation

Domain adaptation is a set of methods that transfer a model trained in a source domain to perform well in a target domain with different data characteristics (e.g., different language, sensors, populations, or operating conditions). It allows an AI trained in one context to be reused reliably in another. Approaches include feature alignment, fine-tuning, and reweighting.

Double Materiality

The concept of Double Materiality highlights the importance to evaluate and disclose publicly both the effects of the company’s activities on the environment and society (environmental and social materiality) as well as the financial impacts of sustainability rules and practices on a company operations and profitability (financial materiality):

• Financial Materiality assesses how ESG factors influence a company’s financial performance, stability, and long-term value.

• Impact Materiality (Environmental and Social) reflects how a company’s activities contribute to or mitigate issues like climate change, resource depletion, human rights, and community well-being.

Embeddings

Embeddings are vector representations of discrete objects (e.g., words, sentences, documents) learned so that semantic or functional similarity corresponds to geometric proximity in the vector space. They turn words and sentences into numbers so that similar ideas are placed close together and can be searched or compared by meaning. Text embeddings support tasks such as semantic search, clustering, and retrieval.

Emissions Trading System

See “Carbon Markets”

ESG - Environment, Social and Governance

ESG stands for Environmental (e.g. energy consumption, water usage), Social (e.g. talent attraction, supply chain management) and Governance (e.g. remuneration policies, board governance). ESG factors form the basis for the different SI approaches.

ESG Integration

ESG Integration refers to the explicit inclusion of ESG risks and opportunities into financial decision-making.

For sustainable investing, the term refers to the integration of ESG risks and opportunities into traditional financial analysis and investment decisions based on a systematic process and appropriate research sources.

For sustainable lending it refers to the integration of ESG risks and opportunities in the loan approval, risk assessment and pricing processes.

Explainability

Explainability is the ability to provide human-understandable reasons for a model’s outputs, such as which inputs influenced a prediction and in what direction. Explainability may be intrinsic (from interpretable models) or post-hoc (from explanation techniques applied after training). It is key for mitigating the black-box nature of large AI models, enabling users to detect errors or bias and ensure accountability.

Interpretability

Interpretability is the degree to which a human can directly understand a model’s structure and internal reasoning. It is often higher for simple, transparent models and lower for complex, high-capacity models, though interpretability can sometimes be improved through constraints and model design.

Fine-tuning

Fine-tuning is the continued training of a pretrained model on task- or domain-specific data to improve performance on a target task. It typically updates some or all model parameters and can be performed with supervised objectives or preference-based objectives.

GEAK

Gebäudeausweis der Kantone (GEAK) is a Swiss tool to measure and indicate the energy consumption of a building. It forms an easy tool to compare the energy efficiency of different buildings both for buyers and investors. www.geak.ch

Generative AI (GenAI)

Generative AI refers to models that learn a data-generating distribution and can produce new samples resembling the training data (e.g., text, images, audio, code). Beyond merely classifying information, GenAI can create new material, like writing a summary or drafting a report. GenAI includes autoregressive language models, diffusion models, variational autoencoders, and generative adversarial networks.

Gradient Boosting Machines

Gradient boosting machines are ensemble methods that build a strong predictor by sequentially adding weak learners (often shallow decision trees) to minimize a differentiable loss function via gradient-based updates in function space.

Greenwashing

Greenwashing refers to the situation where there is a discrepancy between the claim a company makes about its sustainability practices and the reality of its activities, thereby misleading stakeholders (potential investors, business partners, regulators and the wider public) on the depth and extent of its sustainability engagements and practices. Greenwashing can include the use of misleading labels or claims (e.g. using the terms "green", "eco-friendly", etc.), selective disclosure or unprecise language (highlighting only certain information while hiding other, etc.

Hallucination

Hallucination is a failure mode in which a generative model, especially a language model, produces content that is not supported by the provided input, retrieved evidence, or verifiable facts. It can manifest as fabricated details, incorrect assertions, or spurious citations that appear plausible.

Human-in-the-loop

Human-in-the-loop refers to system designs where human judgment is integrated into model training, evaluation, or operation, such as labelling data, providing feedback, reviewing outputs, or approving actions. It is commonly used to improve reliability and safety in high-stakes applications.

Intelligent Process Automation (IPA)

Intelligent process automation combines traditional process automation with AI techniques (e.g., NLP, computer vision, machine learning) to automate workflows that involve unstructured inputs, probabilistic decisions, and exceptions. Unlike purely rule-based automation, IPA can generalize from data and improve with experience.

Interpolation

Interpolation is the ability of a model to make reliable predictions for inputs that lie within the range of the training data. Models are most reliable when working with familiar patterns, not extreme or unseen situations. It contrasts with extrapolation, where the model is asked to predict in regions not well represented during training and performance is typically less reliable.

Isolation forest

Isolation forest is an unsupervised anomaly detection algorithm that isolates observations by randomly selecting features and split values to build partitioning trees. Points that require fewer splits to isolate are assigned higher anomaly scores.

Large Language Model (LLM)

A large language model is a neural network trained on massive text datasets to predict the next token, learning statistical regularities of language. LLMs can be adapted to tasks such as summarization, classification, extraction, and dialogue via prompting, fine-tuning, or alignment methods.

LIME (Local Interpretable Model-agnostic Explanations)

LIME is a post-hoc explanation technique that approximates a complex model locally around a specific prediction using an interpretable surrogate model (often linear). The surrogate’s coefficients are used as a proxy for feature importance near that instance.

Machine learning

Machine learning is a set of methods that learn patterns from data to make predictions, classifications, or decisions without being explicitly programmed with task-specific rules. ML includes supervised, unsupervised, and reinforcement learning, as well as many hybrid approaches.

Model compression

Model compression is the set of techniques used to reduce a model’s memory footprint and computational cost while retaining acceptable performance. Common methods include pruning, quantization, knowledge distillation, and low-rank factorization.

Model drift

Models can become outdated as reality changes. Model drift is a change over time in a model’s performance or behavior caused by changes in data, labels, or underlying relationships. Drift may arise from data drift (input distribution changes) or concept drift (the target relationship changes). It indicates the need for retraining.

Multi-layer perceptron

A multi-layer perceptron is a feedforward neural network composed of fully connected layers with nonlinear activation functions. MLPs are universal function approximators and are widely used for structured/tabular prediction tasks.

Natural Language Processing

Natural language processing is the field of AI concerned with computational methods for understanding, generating, and interacting with human language. It allows computers to read, understand, and write text. It includes tasks such as tokenization, parsing, information extraction, translation, summarization, and dialogue.

Neuromorphic hardware

Neuromorphic hardware refers to computer chips designed to work more like the human brain than traditional computers. Instead of processing data in rigid, step-by-step instructions, these systems are built to handle many small operations simultaneously, activating only when something changes. This can make certain AI tasks much more energy-efficient, especially for pattern recognition or continuous sensing. The technology is still experimental and not widely used today.

Neural network

A neural network is a parameterized model composed of layers of interconnected units that transform inputs into outputs through weights and activations. Inspired by the brain, neural networks learn by adjusting connections based on experience. Neural networks can be used for classification, regression, or generation. They are the basis of deep learning systems.

Parameter

A parameter is a learned value within a model (e.g., a weight or bias) that is adjusted during training to reduce the loss function. The collection of parameters determines the function the model computes from inputs to outputs. Large models include hundreds of billions of parameters.

Paris Agreement

Agreed at COP21 in Paris in 2015, the Paris Agreement's central aim is to strengthen the global response to the threat of climate change by keeping a global temperature rise this century well below 2°C above pre-industrial levels and to pursue efforts to limit the temperature increase even further to 1.5°C.

Parsing
Parsing is the process of analysing structured or semi-structured input to determine its syntaxic structure, relationships, and components according to defined rules or grammars. Parsing helps a system understand how pieces of information fit together - for example, identifying sections, tables, or sentence structure rather instead of simply raw text.

Prompt optimization

Prompt optimization is the practice of iteratively refining an existing prompt without changing its underlying task definition or behavioural intent, in order to maximize performance against explicit objectives: accuracy, consistency, cost, latency, or robustness.

Prompt engineering

Prompt engineering is the practice of designing prompts (instructions, constraints, examples, and context) to steer a language model’s behaviour toward desired outputs. It includes techniques such as few-shot prompting, structured output constraints, and tool-use prompting.

Quantum computing

Quantum computing is a computing approach that uses quantum bits (qubits), which can represent multiple states at the same time, to solve certain types of problems more efficiently than classical computers. Quantum computers use quantum effects to process multiple possible states simultaneously. Most applications are still experimental and not ready for everyday use.

Recurrent Neural Network

A recurrent neural network is a neural architecture for sequential data that maintains a hidden state across time steps, enabling dependence on previous inputs. Variants such as Long Short-Term Memory networks (LSTM) and Gated Recurrent Units (GRU) mitigate vanishing gradients and were widely used before transformer architectures became dominant in many NLP tasks.

Regression

Regression is a family of statistical and machine learning methods for modelling relationships between variables and predicting a continuous target. They answer questions like “if X changes, how much does Y change?”. It includes linear regression and many nonlinear extensions (e.g., generalized linear models, tree-based regression, neural networks).

Reinforcement learning

Reinforcement learning is a machine learning method in which an agent interacts with an environment by taking actions and receiving rewards, with the objective of learning a policy that maximizes expected cumulative reward.

Retrieval-Augmented Generation

Retrieval-augmented generation (RAG) is a system architecture that combines information retrieval with text generation: relevant documents or passages are retrieved from an external knowledge source and provided as context to a generative model. RAG improves factuality and specificity when the retrieved evidence is relevant and of good quality.

Robotic Process Automation

Robotic process automation (RPA) is the automation of repetitive, deterministic tasks in digital systems, often by scripting interactions with user interfaces or APIs. RPA typically follows explicit rules (rule-based system) and is best suited to stable processes with structured inputs.

Rule-based system

A rule-based system is an AI or software system that applies explicitly defined logical rules (e.g., if–then statements) to derive conclusions or actions. Rule-based systems are transparent and predictable but struggle when environments change.

Sampling theory (in model validation)

Sampling theory comprises statistical principles for drawing samples from populations, estimating quantities of interest, and quantifying uncertainty. In model evaluation, it builds the basis for experimental design, confidence intervals, hypothesis tests, and error estimation under finite samples.

Scope 1, 2 and 3 emissions

Scope 1, Scope 2, and Scope 3 emissions are different categories of greenhouse gas (GHG) emissions. These categories were established by the Greenhouse Gas Protocol, launched in 1998.

• Scope 1: All direct GHG emissions, i.e. emissions from sources directly owned or controlled by the company. This includes emissions from burning fuel in company-owned vehicles or from running an industrial production plant.

• Scope 2: Indirect GHG emissions associated with energy purchased by the company. These emissions occur at the facility of another company, where the energy is generated, but are a direct result of the company's energy use.

• Scope 3: Other indirect emissions related to upstream and downstream elements of the value chain, such as the extraction and production of purchased materials and fuels, transport-related activities in vehicles not owned or controlled by the reporting entity, electricity-related activities (e.g. T&D losses) not covered in Scope 2, outsourced activities, waste disposal, etc.

Semantic search

Semantic search retrieves information based on meaning rather than exact keyword matches. It typically represents queries and documents as embeddings and ranking by vector similarity. It is useful when relevant content does not use the exact same vocabulary or phrasing than the query. It is core to core to how AI language applications operate.

Sentiment analysis

Sentiment analysis is a natural language processing task that identifies the expressed attitude in text, such as whether it is positive, negative, or neutral - for example, whether a news article, comment, or report expresses approval, concern, or criticism. More advanced versions also detect emotions, opinions, or the degree of subjectivity.

SHAP (SHapley Additive exPlanations)
SHAP is a set of methods that explain a model’s prediction by estimating how much each input feature contributed to the final output. SHAP explains an AI decision by breaking it down into parts, showing which inputs pushed the result up or down. It tells you how the model made its decision, but not whether the decision reflects real-world cause and effect.

Small language model

A small language model (SLM) is a language model with substantially fewer parameters and lower compute requirements than frontier-scale LLMs. They are often optimized for specific tasks, domains, or deployment constraints. SLMs can provide improved latency and cost efficiency if they are matched appropriately to the task.

Stewardship

Stewardship refers to investors addressing concerns of environmental, social and governance issues by voting on such topics or engaging corporate managers and boards of directors on them. Stewardship is utilised to address business strategy and decisions made by the corporation to reduce risk and enhance sustainable long-term shareholder value.

Supervised learning

Supervised learning is a machine learning method in which a model is trained on labelled examples to learn a mapping from inputs to outputs. It includes classification and regression and relies on the quality and representativeness of labels.

Support Vector Machine
A support vector machine is a supervised learning algorithm that separates data into categories by finding the boundary that best divides them while keeping the largest possible distance to the closest data points. It draws the cleanest possible line (or surface) between groups of data, aiming to keep them as far apart as possible.

Sustainable AI

Sustainable AI refers to the design, development, deployment, and governance of artificial intelligence systems in a manner that ensures their long-term environmental, economic, and societal viability. It encompasses responsible use of resources, including energy use, carbon emissions, and hardware lifecycle. It also includes considerations such as fairness, robustness, transparency, and human oversight.

Sustainable Finance

Sustainable finance refers to any form of financial service with the objective of supporting the transition to a sustainable economy and society by integrating environmental, social and governance factors into business and investment decisions. Such finance aims for the lasting benefit to clients, society at large and the planet.

Synthetic data

Synthetic data is artificially generated data intended to approximate properties of real data for purposes such as privacy protection, augmentation, testing, or simulation. It can be produced via mechanistic simulators, statistical models, or generative models, and should be validated against real data for fidelity and bias. Synthetic data is increasingly used in the training, validation, and evaluation of AI systems, particularly where real-world data is scarce or difficult to obtain. It complements real data rather than fully replacing it.

Taxonomy mapping

Taxonomy mapping is the task of assigning items (documents, activities, entities, or labels) to categories in a predefined taxonomy. It is typically supported by automated classification or extraction techniques. This enables the organization, comparison, and analysis of information by aligning raw data to a common classification framework.

Token, tokenization

Tokenization is the process of breaking down raw input data into smaller units called tokens that can be processed by a machine learning model. Tokens are the fundamental units of representation used by sequence-based models. In language models they correspond to words or subwords, while in other applications they may represent image patches, audio segments, or time-series intervals.

Training

Training is the process through which a model learns by adjusting its internal parameters based on data, repeatedly comparing its predictions to desired outcomes and reducing errors according to an objective function. AI systems learn from examples: they makes a guess, check how wrong it was, adjust themselves, and repeat this many times until performance improves.

Transformer

A transformer is a neural network architecture for sequence modelling that relies on attention mechanisms (especially self-attention) to represent relationships between tokens. Transformers enable parallel processing of sequences during training and are the foundation of most modern large language models.

Unsupervised learning

Unsupervised learning is a machine learning method in which models learn structure from unlabelled data, such as clusters, low-dimensional representations, or density estimates. Common tasks include clustering, anomaly detection, and dimensionality reduction.

Vector database

A vector database is a data store optimized for indexing and querying high-dimensional vectors (embeddings) using similarity search, often via approximate nearest neighbour algorithms. It is commonly used to support semantic search and retrieval-augmented generation systems.